The mutex locking scheme in FaultManager was recently revised in SERVER-109338.

Those changes introduced a bug. runMongosServer can indirectly call FaultManager::_init() concurrently with the previously started FTDC thread indirectly calling FaultManager::appendDescription(...). These two functions contend for the data member std::vector<std::unique_ptr<HealthObserver>> FaultManager::_observers.

This race was identified in code review, but then the commit that addressed the review comment did not actually fix the race condition.